Ordinal contrastive learning for imputing missing data with progressive labels

Abstract

A method of training a generative model for missing data imputation comprises: acquiring a first dataset; embedding, using an encoder included in the generative model, each of a plurality of data included in the first dataset into an embedding space; reconstructing, using a decoder included in the generative model, each of the plurality of data included in the first dataset based on a result of the embedding; and training the encoder based on the result of the embedding. A method of missing data imputation comprises: acquiring a missing dataset; acquiring a target modality condition; inputting the missing dataset and the target modality condition into a generative model; and imputing missing data based on an output value of the generative model.

What is claimed is: 
     
         1 . A method of training a generative model for missing data imputation, the method comprising:
 acquiring, by an electronic device, a first dataset;   embedding, by the electronic device using an encoder included in the generative model, each of a plurality of data included in the first dataset into an embedding space;   reconstructing, by the electronic device using a decoder included in the generative model, each of the plurality of data included in the first dataset based on a result of the embedding; and   training, by the electronic device, the encoder based on the result of the embedding,   wherein the first dataset comprises ordered data across a plurality of modality types,   wherein the first dataset is a missing dataset in which a portion of data is missing, and   wherein training the encoder comprises training the encoder such that the positions of the plurality of data in the embedding space are determined based on their order in the first dataset.   
     
     
         2 . The method of  claim 1 , wherein the decoder reconstructs each of the plurality of data included in the first dataset by receiving a source modality condition together with the result of the embedding. 
     
     
         3 . The method of  claim 1 , wherein the first dataset comprises data acquired from a plurality of sources, and wherein training the encoder comprises training the encoder such that data from an identical source among the data included in the first dataset are embedded in similar positions in the embedding space. 
     
     
         4 . The method of  claim 1 , wherein training the encoder comprises training the encoder such that the encoder does not consider modality types of the plurality of data included in the first dataset. 
     
     
         5 . The method of  claim 1 , further comprising training, by the electronic device, the decoder based on the reconstructed result, wherein training the decoder comprises training the decoder such that the decoder reconstructs the original input data. 
     
     
         6 . An electronic device comprising:
 an input device configured to acquire a first dataset;   a processor configured to:
 embed, using an encoder included in a generative model, each of a plurality of data included in the first dataset into an embedding space; 
 reconstruct, using a decoder included in the generative model, each of the plurality of data included in the first dataset based on a result of the embedding; and 
 train the encoder based on the result of the embedding; 
   a storage device configured to store the generative model,   wherein the first dataset comprises ordered data across a plurality of modality types,   wherein the first dataset is a missing dataset in which a portion of data is missing,   wherein training the encoder comprises training the encoder such that the positions of the plurality of data in the embedding space are determined based on their order in the first dataset.   
     
     
         7 . The electronic device of  claim 6 , wherein the decoder reconstructs each of the plurality of data included in the first dataset by receiving a source modality condition together with the result of the embedding. 
     
     
         8 . The electronic device of  claim 6 , wherein the first dataset comprises data acquired from a plurality of sources, and wherein training the encoder comprises training the encoder such that data from an identical source among the data included in the first dataset are embedded in similar positions in the embedding space. 
     
     
         9 . The electronic device of  claim 6 , wherein training the encoder comprises training the encoder such that the encoder does not consider modality types of the plurality of data included in the first dataset. 
     
     
         10 . The electronic device of  claim 6 , wherein the processor is further configured to train the decoder based on the reconstructed result, wherein training the decoder comprises training the decoder such that the decoder reconstructs the original input data. 
     
     
         11 . A method of missing data imputation, the method comprising:
 acquiring, by an electronic device, a missing dataset;   acquiring, by the electronic device, a target modality condition;   inputting, by the electronic device, the missing dataset and the target modality condition into a generative model; and   imputing, by the electronic device, missing data based on an output value of the generative model,   wherein the generative model is trained by the method of  claim 1 .   
     
     
         12 . The method of  claim 11 , wherein the missing dataset comprises medical imaging data.